Remote Sensing of Environment 195 (2017) 1–12

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Effects of the spatial configuration of trees on urban heat mitigation: A

comparative study
Weiqi Zhou a,b,⁎, Jia Wang a,b, Mary L. Cadenasso c
a
State Key Laboratory of Urban and Regional Ecology, Research Center for Eco-Environmental Sciences, Chinese Academy of Sciences, No. 18 Shuangqing Road, Beijing 100085, China
b
University of Chinese Academy of Sciences, No. 19A Yuquan Road, Beijing 100049, China
c
Department of Plant Sciences, University of California, Davis, One Shields Ave., Davis, CA 95616, USA

a r t i c l e i n f o a b s t r a c t

Article history: Urban greenspace has significant cooling effects on urban heat. Recent studies investigating the effects of spatial
Received 2 June 2016 configuration of greenspace show significant, but inconsistent results, including both positive and negative ef-
Received in revised form 22 March 2017 fects. To investigate the causes of this inconsistency, we compared Baltimore, MD and Sacramento, CA, USA,
Accepted 30 March 2017
two cities with very different climatic conditions. We quantified and compared the relationships between the
Available online xxxx
spatial configuration of trees and land surface temperature (LST) using different statistical approaches, and con-
Keywords:
ducted the analyses using spatial units of different sizes, based on trees mapped from 1 m high resolution imag-
Urban tree canopy ery. We found: (1) trees' cooling efficiency was higher in Baltimore than in hotter and drier Sacramento.
Spatial configuration Additionally, percent cover of trees was more important than their spatial configuration in predicting LST in Bal-
Urban heat mitigation timore, but the opposite was found in Sacramento. (2) Spatial configuration of trees affects LST more in Sacra-
Urban ecology mento than in Baltimore, and the effects of spatial configuration of trees on LST varied greatly in terms of
Baltimore magnitude, significance, and even direction, between the two cities. Notably, mean patch size had significantly
Sacramento positive effects on LST in Baltimore, but negative effects in Sacramento. In contrast, edge density had negative ef-
fects on LST in Baltimore, but positive effects in Sacramento. (3) Different statistical approaches resulted in dra-
matic changes in the relationships between LST and configuration metrics. Our results underscore the necessity
of controlling the effects of percent cover of trees, when quantifying the effects of spatial configuration of trees on
LST. (4) Spatial autocorrelation may influence relationships between landscape metrics and LST, particularly
when the unit of analysis is relatively small. (5) The relationships between spatial configuration metrics and
LST are stronger with an increase of the size of the analytical unit. This study can enhance our understanding
of the effects of spatial configuration of greenspace on urban heat island (UHI). It also provides important insights
to urban planners and natural resource managers on how to mitigate the impact of urbanization on UHI through
urban design and vegetation management.
© 2017 Elsevier Inc. All rights reserved.

1. Introduction mortality and morbidity in cities worldwide (Fouillet et al., 2006;

Harlan and Ruddell, 2011). Consequently, how to mitigate and adapt
Urban heat island (UHI) describes the phenomenon by which urban to the UHI has become a major research focus in urban climatology
areas are warmer than surrounding non-urban areas (Voogt and Oke, and urban ecology (Arnfield, 2003; Sun and Chen, 2017; Weng, 2009;
2003). Increased temperatures due to the UHI effect may increase Zhou et al., 2011).
water consumption and energy use in urban areas (Santamouris et al., Considerable research has demonstrated the significant cooling ef-
2015; Wan et al., 2012), alter species composition and distribution fects of urban greenspace on urban heat (Fan et al., 2015; Jenerette et
(Niemelä, 1999; White et al., 2002), and lead to an increase in the pro- al., 2007; Kong et al., 2014; Li et al., 2016; Ma et al., 2010; Weng et al.,
duction of ground level ozone which has direct consequences for 2004; Zhou et al., 2011). Increasing the percent cover of greenspace
human health (Akbari et al., 2001; Akbari et al., 1996). In addition, ex- can greatly reduce ambient air temperatures and land surface tempera-
cess heat affects the comfort of urban dwellers and leads to greater tures (Bowler et al., 2010; Connors et al., 2013; Fan et al., 2015; Li et al.,
health risks (Poumadere et al., 2005). In fact, extreme heat increases 2012; Weng et al., 2004; Zhou et al., 2011; Zhou et al., 2014). In addition,
the spatial configuration (or arrangement) of greenspace, can also have
significant effects on land surface temperature (LST) (Chen et al., 2014;
⁎ Corresponding author at: State Key Laboratory of Urban and Regional Ecology,
Research Center for Eco-environmental Sciences, Chinese Academy of Sciences, No. 18
Fan et al., 2015; Kong et al., 2014; Li et al., 2013b; Li et al., 2012;
Shuangqing Road, Beijing 100085, China. Maimaitiyiming et al., 2014; Myint et al., 2015; Zhou et al., 2011). Be-
E-mail address: wzhou@rcees.ac.cn (W. Zhou). cause cities have limited space for greening, managers and decision-

http://dx.doi.org/10.1016/j.rse.2017.03.043
0034-4257/© 2017 Elsevier Inc. All rights reserved.
2 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12

makers would benefit from knowing how to optimize the spatial config- precipitation is approximately 1070 mm. Sacramento is the capital city
uration of greenspace to further alleviate urban heat stress (Huang et al., of California. It has a total area of 259 km2, and total population of about
2011; Li et al., 2016; Myint et al., 2015; Zhou et al., 2011). 0.48 million in 2014. Located at the confluence of the Sacramento and
We know that simply increasing the percent cover of greenspace American rivers, its annual average temperature is 16.2 °C and average
leads to a reduction of temperatures; this relationship is very consistent. precipitation is approximately 450 mm. The similarity in the sizes of
What is less known, however, is the effects of the spatial configuration total population and area, but the contrast in climatic conditions and bi-
of that greenspace on urban temperatures. Research results are, in omes, make the two cities ideal for the comparisons conducted in this
some cases, contradictory. For example, greater patch density of research.
greenspace reduced LST in studies conducted in Shenzhen (Li et al.,
2010) and Shanghai, China (Li et al., 2011), Baltimore, USA (Zhou et 2.2. Data
al., 2011), and Berlin, Germany (Dugord et al., 2014), but was associated
with increased LST in Beijing, China (Li et al., 2013b; Li et al., 2012). Sim- 2.2.1. Land surface temperature
ilarly, edge density of greenspace was found to be negatively correlated The LST data were derived from the thermal infrared (TIR) band
to LST in many cities (Dugord et al., 2014; Li et al., 2011; Li et al., 2014; (10.40–12.50 μm) of two Landsat-5 Thematic Mapper (TM) images
Maimaitiyiming et al., 2014; Rhee et al., 2014; Zhang et al., 2009; Zhou with a spatial resolution of 120 m (Fig. 1 BLST, SLST). The TM data for Bal-
et al., 2011), but positively correlated in others (Li et al., 2013b; Wu et timore and Sacramento were acquired on August 11, 2007 (row 33/path
al., 2014). This inconsistency prevents the application of results to 15), and August 14, 2010 (row 33/path 44), respectively. LST was de-
urban greenspace planning and management (Li et al., 2013b). rived for different years in order to coincide with the years the land
The reasons for this inconsistency remain largely unaddressed. It cover for the two cities was collected – Baltimore in 2007 and Sacra-
may be because these studies have been conducted 1) in cities with con- mento in 2010.
trasting climatic conditions; 2) using a variety of statistical analysis (Fan We first calculated the top-of-atmospheric (TOA) radiance based on
et al., 2015; Kong et al., 2014; Li et al., 2013b; Li et al., 2012; Myint et al., the digital number (DN) of the TM TIR band (Chander and Markham,
2015; Zhou et al., 2011); 3) based on maps from image data with spatial 2003; Landsat Project Science Office, 2009). We then calculated the sur-
resolution ranging from sub-meter to1000 m (Li et al., 2013b; Rhee et face-leaving radiance from TOA radiance by removing the effects of the
al., 2014; Wu et al., 2014; Zhou et al., 2011); and 4) using a variety of an- atmosphere in the thermal region (Asgarian et al., 2015; Barsi et al.,
alytical units with different sizes such as grids or pixels (Peng et al., 2005; Sobrino et al., 2004; Yuan and Bauer, 2007; Zhou et al., 2014). Fi-
2016; Rhee et al., 2014), city blocks (Dugord et al., 2014), sub-districts nally, LST was calculated from surface-leaving radiance using the Plank
(Li et al., 2013b), or self-defined polygons (Zhou et al., 2011). Does spa- function (Chander and Markham, 2003; Chander et al., 2009).
tial configuration of greenspace affect temperatures differently in cities
with different climatic conditions? Or, is this inconsistency due to the 2.2.2. Spatial pattern of tree canopy
varied statistical approaches applied, or different units of analysis, or We mapped the urban tree canopy based on 1-m resolution imagery
different resolutions of data to map greenspace? from the National Agriculture Imagery Program (NAIP), using an object-
Here, we address these questions by conducting a comparison study based classification approach (MacFaden et al., 2012; Zhou and Troy,
of Baltimore, MD and Sacramento, CA, USA, two cities with very differ- 2008). The imagery is 4-band color-infrared, with radiometric depth
ent climatic conditions. We quantified and compared the relationships of 8 bits. Ancillary data, such as light detecting and ranging (Lidar)
between spatial configuration of trees and LST using different statistical data and building footprint layers, were used to aid in classification.
approaches, and conducted the analyses at sampling units of different Six classes were included in the classification map: trees (i.e., tree cano-
sizes. We mapped tree canopies using 1 m resolution imagery. This de- py), grasses, pavement, buildings, water and bare soil (Fig. 1 BTC, STC).
cision was based on the work of Li et al. (2013b), Zhou et al. (2014) and The accuracies of the land cover classifications were assessed by visually
Zhou et al. (2016), which suggested that the spatial resolution of image referencing to sub-meter high-resolution imagery using protocol devel-
data used to map greenspace influenced the statistical relationships be- oped in Zhou and Troy (2008). The overall accuracies of the classifica-
tween spatial configuration of greenspace and LST, and that high spatial tions were 95.7% for Baltimore and 93.6% for Sacramento. The user's
resolution image data are more appropriate in such analysis. Results and producer's accuracy of trees for Baltimore were 97.3% and 97.5%,
from the present study can enhance the understanding of the effects and 98.2% and 96.7% for Sacramento.
of spatial configuration of greenspace on UHI. In addition, important in- There are numerous metrics that can be used to measure and de-
sights can be provided to urban planners and natural resource managers scribe spatial patterns of land cover features (Gustafson, 1998;
on how to mitigate the impact of urbanization on UHI through urban McGarigal et al., 2002). Here, we chose 5 landscape metrics to measure
design and vegetation management. the spatial pattern of urban trees, including one composition metric:
percent cover of trees (PTree), and four configuration metrics: (1)
2. Methods mean patch size (AREA_MN), (2) edge density (ED), (3) mean patch
shape index (SHAPE_MN), and (4) largest patch index (LPI) (Table 1).
2.1. Study area These metrics represent the primary characteristics describing the spa-
tial pattern of trees, including the abundance of trees, size and shape of
The research focuses on two cities with contrasting climatic condi- patches, edge density, and fragmentation. These metrics were chosen
tions, Baltimore, Maryland, USA, and Sacramento, California, USA. Balti- based on the following considerations: (1) importance in both theory
more is a temperate coastal city characterized by hot and humid and practice (Lee et al., 2009; Li and Wu, 2004; Peng et al., 2010; Zhou
summers (Brazel et al., 2000), while Sacramento has a Mediterranean et al., 2011), (2) easily calculated and interpretable (Li et al., 2012;
climate characterized by hot, but dry summers. Baltimore is built in a Zhou et al., 2011), and (3) minimal redundancy (Riitters et al., 1995;
biome dominated by temperate broadleaf and mixed forest, whereas Li and Wu, 2004; Zhou et al., 2011). These metrics were calculated in
Sacramento belongs to a biome dominated by grassland, with riparian ArcGIS™ 10.1.
forests only along the streams and shrub and woodlands that do not
occur until in the sierra foothills and higher elevation (Imhoff et al., 2.3. Statistics analysis
2010).
Baltimore is the largest city in Maryland, with a total area of 239 km2 We investigated the relationships among spatial patterns of tree
and total population of approximately 0.62 million in 2014. Close to the canopy and LST at multiple scales, that is, using different sizes of analyt-
Chesapeake Bay, its annual average temperature is 12.6 °C, and average ical units. Specifically, 5 sizes of analytical unit were used: 1) 1 × 1 pixel
 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12 3

Fig. 1. The spatial distribution of tree canopy and land surface temperatures in Baltimore (panels BTC and BLST) and Sacramento (Panels STC and SLST).

(or a grid cell of 120 m × 120 m, the same as the pixel size of the Landsat statistical analyses. The predictor variables were the percent cover of
TM thermal band), 2) 3 × 3 pixels (360 m × 360 m), 3) 5 × 5 pixels tree canopies, and the four landscape metrics (Table 1). Table A1
(600 m × 600 m), 4) 7 × 7 pixels (840 m × 840 m), and 5) 9 × 9 pixels shows the mean and standard deviation of LST and landscape metrics.
(1080 m × 1080 m) (Liu and Weng, 2009). For each analytical unit (i.e., A Pearson correlation matrix was first developed to examine the cor-
a grid cell), we calculated the mean LST as the response variable for relations between LST and the spatial pattern metrics of trees. We then
4 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12

Table 1
Landscape metrics used in this study, after McGarigal et al. (2002).

Categories Landscape metrics Description Equation Citations

(abbreviation) (unit)
n 
Composition Percent cover of tree canopy Proportion of tree canopy area within an analysis unit. ∑i¼1 ai
100 (Li et al., 2014; Zhou et al., 2011)
A
(PTree)
(%)
n
Configuration Mean patch size (AREA_MN) The average area of tree canopy patches within an analysis unit. ∑i¼1 ai (Kong et al., 2014; Zhang et al., 2009)
N
2
(m )
0:25 pi
Mean patch shape index The average shape index of tree canopy patches within an n
∑i¼1 p
A
ffi (Li et al., 2012; Peng et al., 2010)
(SHAPE_MN) analysis unit. N

Edge density (ED) The total perimeter of tree canopy patches per km2 within an n
∑i¼1 pi
10000 (Connors et al., 2013; Maimaitiyiming et
A
analysis unit. al., 2014)
(m/ha)
maxai 
Largest patch index (LPI) The proportion of the largest tree canopy patch within an A 100 (Rhee et al., 2014; Zhou et al., 2011)
analysis unit. (%)

ai area of tree canopy patch i; pi perimeter of tree canopy patch i; A total area of analysis unit; N number of patches of tree canopy.

conducted a partial correlation analysis to investigate the relationships We used the Lagrange Multiplier statistics to compare the two
between LST and the configuration metrics, by controlling for the effect modeling approaches, and found that the spatial error model better fit
of the percent cover of trees. Controlling for the effect tree canopy per- the data in this study. The regressions were then run using the spatial
cent is necessary because the configuration metrics were highly corre- error model, and a maximum likelihood method. The R2 values were
lated to percent cover of trees, and therefore the Pearson correlation calculated as detailed in Lichstein et al. (2002), which were comparable
analysis may obtain spurious relationships between LST and configura- with those from the OLS regression model. The regressions were run in
tion metrics. GeoDa 1.6.7 and spdep package of R (Version 2.12.1; R Development
We then used ordinary least squares (OLS) multiple linear regres- Core Team, 2011).
sion model and spatial autoregression (SAR) model to examine the ef- Variance partitioning was used to quantify the relative variations in
fects of the spatial pattern of trees on LST. We used standardized LST explained by: the percent cover of trees and the configuration met-
coefficients (beta weights) to evaluate the relative importance of per- rics. The variation of LST was divided into four fractions: (1) unique ef-
cent cover and configuration metrics on predicting LST (Weng et al., fects of percent cover of trees, (2) unique effects of configuration
2006; Yan et al., 2014; Zhou et al., 2011), and variance partitioning to metrics, (3) joint effects of percent cover of trees and configuration met-
quantify the explanatory power of the predictors (Anderson and rics, and (4) unexplained. Variance partitioning was conducted follow-
Gribble, 1998; Li et al., 2013a; Li et al., 2012). ing the procedure detailed in Anderson and Gribble (1998) and in
The OLS regression model is the most commonly used statistical Heikkinen et al. (2005), using the spdep package (Anselin, 2005b) of R
analysis, with the assumption that the error terms are independent. (Version 2.12.1; R. Team, 2011).
The primary analyses showed that significant spatial autocorrelation
(P b 0.01) occurred in the residuals of the OLS model. Consequently, spa- 3. Results
tial autoregression models that integrate spatial autocorrelation into
modeling were more appropriate to investigate the relationships be- 3.1. The spatial distribution of trees and LST in the two cities
tween LST and spatial patterns of trees (Li et al., 2012). We also included
the OLS regression model for comparison purposes, as many studies in The percent cover of trees, as well as the spatial configuration, dif-
the literature use such analyses. Below, we briefly describe the spatial fered greatly between the two cities (Fig. 1 BTC, STC). Approximately
autoregression models and variance partitioning. More details can be 27.1% of the land in Baltimore was covered by trees, but only 16.7% in
found in Li et al. (2012). Sacramento. Compared to Sacramento, trees in Baltimore are more clus-
With SAR, the neighborhood relationship of the response variable is tered, especially in the northwest region of the city (Fig. 1 BTC). For both
explicitly measured by a (n x n) matrix of spatial weights, which is inte- cities, percent cover of trees varied greatly in space. Taking the analyti-
grated into the standard multiple linear regression to account for spatial cal unit of 600 × 600 m as an example, percent cover of trees in grid cells
autocorrelation (Anselin, 2005a). The spatial autocorrelation can be varied from 0.50% to 92.62% across Baltimore, with a standard deviation
modeled in two ways: a spatial lag model and a spatial error model of 18.73%. In Sacramento, percent cover ranged from 0 to 58.68%, with a
(Anselin, 2005a). The spatial lag model assumes that the spatial standard deviation of 12.12% (Table A1). The mean patch size of trees in
autoregressive occurs only in the response variable. The form of the spa- Baltimore was 599.6 m2, much greater than that of 73.80 m2 in Sacra-
tial lag model is: mento. In contrast, the patch density and edge density of trees in Sacra-
mento were much higher than that of Baltimore (2227/km2 versus 399/
y ¼ ρWy þ βX þ ε ð1Þ km2 for patch density and 819.85 m/ha versus 422.31 m/ha for edge
density), suggesting that tree cover was more fragmented in Sacramen-
where Wy is a (n × 1) vector of the spatially lagged response variable, ρ to. The mean shape index was similar in the two cities (1.32 in Baltimore
is a spatial autoregressive coefficient, X is a (n x k) vector of explanatory and 1.39 in Sacramento), suggesting that the complexity of the tree
variables, β is a (k × 1) vector of regression coefficients, and ε is a (n × 1) patches is similar.
vector of independently distributed errors. Land surface temperatures varied greatly in space for both cities
In contrast, the spatial error model assumes the spatial effects that (Fig. 1 SLST, SLST). LST in Baltimore ranged from 8.93 °C to 50.99 °C,
are not fully explained by the explanatory variables occurs in the error with a mean of 33.37 °C and standard deviation of 4.69 °C, while it
terms, and therefore, is expressed as: ranged from 12.17 °C to 50.11 °C, with a mean of 35.60 °C and standard
deviation of 3.25 °C in Sacramento (Table A1). For both cities, LST was
y ¼ βX þ λWμ þ ε ð2Þ significantly autocorrelated in space, as indicated by Moran's I (Balti-
more: Moran's I = 0.88, p b 0.01; Sacramento: Moran's I = 0.72, p b
where Wμ is a (n × 1) vector of spatially lagged errors, and λ is a spatial 0.01). LST tended to be higher in locations with less tree canopy cover-
autoregressive coefficient. age (Fig. 1 BTC, BTC, SLST, SLST).
 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12 5

3.2. Effect of spatial patterns of trees on LST: difference between cities and effects of percent cover of trees, the correlation between mean patch
across analytical scales size (AREA_MN) and LST changed from negative to positive when the
analytical unit was less than or equal to 360 m, and then to no longer
3.2.1. Effects of percent cover of trees on LST significant in Baltimore. Similarly, edge density (ED) was no longer sig-
The percent cover of trees was significantly negatively correlated nificantly correlated to LST at the analytical unit N120 m on a side. In
with LST, across all analytical scales, for both cities, suggesting LST de- Sacramento, however, AREA_MN still had a relatively strong negative
creased with the increase of percent cover of trees (Table 2; Fig. A1). relationship with LST across all scales, but the relationships between
The Pearson correlation analysis showed that percent cover of tree can- ED and LST changed from negative to positive. SHAPE_MN remained
opy had the strongest correlation with LST among the 5 metrics. For significantly correlated with LST in Baltimore, but not in Sacramento.
both cities, the strength of the correlations between LST and percent LPI remained significantly correlated with LST for both cities. However,
cover of trees, as indicated by the correlation coefficients, increased these correlations changed from negative to positive in Baltimore, when
with the increase of the size of the analytical unit. The correlations be- controlling for the effects of percent cover of trees (Table 2). For all 4
tween LST and percent cover of trees, however, were generally stronger configuration metrics, the partial correlations were stronger in Sacra-
in Baltimore than in Sacramento across all 5 analytical scales, suggesting mento than in Baltimore, in contrast to the Pearson correlations.
that percent cover of trees might explain more variations of LST in
milder coastal regions compared to hotter and drier ones. 3.2.3. Relative importance of amount and configuration of trees on LST
Results from the OLS multiple linear regressions showed that in Balti-
3.2.2. Effects of spatial configuration of trees on LST more, percent cover of trees (PTree) had significantly negative effects on
The Pearson correlation analysis showed that all 4 metrics of tree LST, across the 5 analytical scales (Table 3). In addition, PTree was the
configuration were significantly, negatively correlated with LST, across most important predictor of LST, playing a much more important role in
all analytical scales, for both cities (Table 2). Similar to percent cover predicting LST than the other spatial configuration variables, as suggested
of trees, the strength of the correlations between LST and the 4 configu- by the standard coefficients (Table 3). None of the configuration variables
ration metrics also generally increased with the increase of the size of were significant at any analytical scale. Among the 4 configuration
the analytical unit; the correlations between LST and the 4 configuration metrics, shape index (SHAPE_MN) played a relatively important role in
metrics were stronger in Baltimore than in Sacramento. Among the 4 predicting LST, and had a negative effect (Table 3). Results from the vari-
configuration metrics, the largest patch index had relatively strong cor- ation partitioning also indicated that percent cover of trees played a more
relations with LST. important role than that of configuration of trees (Fig. 2).
After controlling for the effects of percent cover of trees, the correla- In Sacramento, however, the relative importance of percent cover of
tions (i.e., partial correlations) between configuration metrics and LST trees (PTree) and spatial configuration differed greatly from that of Bal-
changed greatly, as indicated by the results from the partial correlation timore. PTree became no longer significantly related to LST at the ana-
analysis (Table 2). These changes included the following: 1) the lytical units having length scales of 840 m and 1080 m. In contrast,
strength of partial correlations, measured by the partial correlation co- mean patch size (AREA_MN) was significant at all 5 analytical units,
efficients, greatly decreased, compared with their corresponding Pear- and shape index (SHAPE_MN) and edge density (ED) were significant
son correlation coefficients; 2) some of the configuration metrics were at all scales except for 360 m. In addition, configuration metrics became
no longer significantly correlated to LST; and 3) more notably, the rela- more important in predicting LST, with AREA_MN being the most im-
tionships between some of the configuration metrics and LST changed portant predictor of LST for analytical units larger than 120 m on a
from negative to positive. side (Table 3). Results from the variation partitioning also indicated
These changes in the relationships between LST and configuration that configuration of trees played a more important role than that of
metrics, however, varied dramatically in the two cities, in terms of mag- percent cover of trees (Fig. 2).
nitude, significance, and direction. Specifically, after controlling for the Overall, results from the spatial error models were similar to those of the
OLS regression models (Table 4). This was particularly true when the ana-
lytical units were relatively large. For example, when the analytical unit
Table 2
Correlation coefficients between LST and landscape metrics. The italic and bold rows are was greater than or equal to 600 m on a side, the coefficients of the predic-
for partial correlation analysis, where for configuration metrics, the control variable was tors, and the R2 values were similar between OLS models and spatial error
percent cover of tree, and for percent cover of tree, the control variables were the config- models. However, it should be noted that at the analytical length scale of
uration metrics. 120 m, the absolute values of coefficients from the spatial error models
City Scale PTree AREA_MN SHAPE_MN ED LPI were much smaller than those from the OLS regression models, suggesting
Baltimore 120 m −0.830⁎⁎ −0.561⁎⁎ −0.418⁎⁎ −0.559⁎⁎ −0.782⁎⁎
the importance of considering spatial autocorrelation at finer scales.
−0.260⁎⁎ 0.054⁎⁎ 0.003 −0.033⁎⁎ 0.036⁎⁎ For both OLS and SAR, results from the standard coefficients and var-
360 m −0.904⁎⁎ −0.453⁎⁎ −0.604⁎⁎ −0.555⁎⁎ −0.805⁎⁎ iance partitioning showed that among the five metrics, PTree was the
−0.502** 0.059⁎ −0.143⁎⁎ −0.007 0.076⁎⁎ most important predictor of LST in Baltimore. In Sacramento, however,
600 m −0.926⁎⁎ −0.478⁎⁎ −0.752⁎⁎ −0.578⁎⁎ −0.796⁎⁎
configuration metrics, such as AREA_MN, were better predictors of LST
−0.499⁎⁎ 0.085 −0.308⁎⁎ −0.011 0.123⁎⁎
840 m −0.937⁎⁎ −0.687⁎⁎ −0.764⁎⁎ −0.604⁎⁎ −0.779⁎⁎ than PTree, when the size of analytical unit was N120 m (for OLS) or
−0.574⁎⁎ 0.119 −0.363⁎⁎ −0.053 0.185⁎⁎ 360 m (for SAR).
1080 −0.948⁎⁎ −0.535⁎⁎ −0.767⁎⁎ −0.618⁎⁎ −0.778⁎⁎
m −0.562⁎⁎ 0.149 −0.390⁎⁎ −0.116 0.224⁎ 4. Discussion
Sacramento 120 m −0.640⁎⁎ −0.354⁎⁎ −0.361⁎⁎ −0.464⁎⁎ −0.602⁎⁎
−0.234⁎⁎ −0.147⁎⁎ −0.115⁎⁎ 0.197⁎⁎ −0.157⁎⁎
360 m −0.723⁎⁎ −0.704⁎⁎ −0.332⁎⁎ −0.525⁎⁎ −0.611⁎⁎ 4.1. The effects of tree cover and its spatial configuration on LST: relative im-
−0.134⁎⁎ −0.432⁎⁎ −0.051⁎ 0.309⁎⁎ −0.219⁎⁎ portance varied greatly between cities in different climatic zones
600 m −0.768⁎⁎ −0.750⁎⁎ −0.345⁎⁎ −0.564⁎⁎ −0.588⁎⁎
−0.087⁎ −0.475⁎⁎ 0.041 0.341⁎⁎ −0.238⁎⁎
Percent cover of trees had similar effects on LST for both cities de-
840 m −0.811⁎⁎ −0.788⁎⁎ −0.545⁎⁎ −0.609⁎⁎ −0.609⁎⁎
−0.105 −0.529⁎⁎ 0.025 0.375⁎⁎ −0.253⁎⁎ spite the different climatic conditions of these cities. These results are
1080 −0.819⁎⁎ −0.822⁎⁎ −0.578⁎⁎ −0.610⁎⁎ −0.589⁎⁎ similar to findings from previous studies (Li et al., 2011; Li et al.,
m 0.047 −0.565⁎⁎ −0.022 0.410⁎⁎ −0.259⁎⁎ 2013b; Weng et al., 2004; Zhou et al., 2011). Increasing the percent
⁎⁎ P b 0.01. cover of trees can significantly decrease LST for both cities. However,
⁎ P b 0.05 (2-tailed). the efficiency in cooling, defined as the decrease in degrees of LST
6 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12

Table 3
Results from the OLS multiple linear regressions and the diagnostics for spatial dependence. The bold and italic rows are standardized coefficients.

City Scale PTree AREA_MN SHAPE_MN ED LPI R2 Moran's I AIC

Baltimore 120 m −0.154⁎⁎ 7.879E−05⁎⁎ 0.024 1.772E−06 0.007 0.690 0.602 63,539.400
−0.911 0.043 0.002 0.014 0.042
360 m −0.195⁎⁎ 5.860E−05⁎⁎ −2.071⁎⁎ 1.359E−05⁎⁎ 0.024⁎⁎ 0.824 0.460 5478.030
−1.046 0.060 −0.091 0.085 0.126
600 m −0.168⁎⁎ 1.291E−04⁎⁎ −8.697⁎⁎ 1.225E−05⁎ 0.001 0.877 0.427 1619.540
−0.883 0.099 −0.203 0.073 0.005
840 m −0.179⁎⁎ 2.168E−04 −10.290⁎⁎ 1.692E−05⁎ 0.003 0.897 0.372 708.088
−0.934 0.088 −0.202 0.097 0.017
1080 m −0.168⁎⁎ 1.080E−04 −11.524⁎⁎ 9.709E−06 -0.003 0.915 0.424 365.764
−0.865 0.064 −0.199 0.054 −0.014
Sacramento 120 m −0.160⁎⁎ −0.001⁎⁎ −0.993⁎⁎ 1.435E−05⁎⁎ -0.002 0.449 0.653 75,885.800
−0.817 −0.039 −0.138 0.304 −0.006
360 m −0.083⁎⁎ −0.014⁎⁎ 0.309 −1.991E−06 0.008 0.612 0.393 6578.530
−0.419 −0.434 0.018 −0.041 0.019
600 m −0.047⁎ −0.019⁎⁎ 3.649⁎⁎ −1.137E−05⁎⁎ 0.016 0.696 0.323 1973.260
−0.245 −0.596 0.142 −0.231 0.033
840 m −0.051 −0.023⁎⁎ 7.498⁎⁎ −1.450E−05⁎ 0.068⁎ 0.770 0.369 823.595
−0.266 −0.700 0.203 −0.298 0.114
1080 m 0.025 −0.032⁎⁎ 8.182⁎⁎ −2.671E−05⁎⁎ 0.028 0.796 0.318 413.287
0.131 −0.876 0.212 −0.546 0.046
⁎⁎ P b 0.01.
⁎ P b 0.05 (2-tailed).

with every 1% increase in tree cover (Buyantuyev and Wu, 2010; 2015), and its cooling effectiveness is likely to be more affected by dif-
Hamada and Ohta, 2010; Li et al., 2013b; Peng et al., 2016; Xie et al., ferent management practices such as irrigation.
2013), was higher in Baltimore than in Sacramento at all five scales of The cooling efficiency of urban trees can be affected by many factors
analytical unit (Table 5). The results remained the same even after con- such as tree species, spatial configuration of trees, and management
sidering the effects of spatial configuration, except for the analysis at the practices because, for example, transpiration rates of urban trees vary
scale of 120 m (Tables 3 & 4). These results contrast with previous work greatly by species (Pataki et al., 2011; Wang et al., 2011), and are affect-
conducted within southern California that showed more effective ed by climatic factors such as air temperature, total radiation, vapor
cooling by vegetation in hotter and drier desert regions compared to pressure deficit, and ambient pollutants such as ozone (Wang et al.,
milder coastal ones (Tayyebi and Darrel Jenerette, 2016). However, it 2011). These contrasting results warrant further research on the cooling
should be noted that Tayyebi and Darrel Jenerette (2016) used the nor- effectiveness of vegetation/trees that requires field work on species
malized difference vegetation index (NDVI) to measure the abundance identity, species transpiration rates, vegetation management such as ir-
of vegetation, which includes both trees and grass/lawns. But here we rigation, and more detailed climate records (McCarthy et al., 2011;
used the percent cover of trees. Previous findings have shown that Pataki et al., 2011; Polsky et al., 2014; Zhou et al., 2008).
grass is less effective than tree canopy for LST cooling (Myint et al., Effects of spatial configuration of tree cover on LST, however, varied
greatly in the two cities, in terms of magnitude, significance, and even
direction of effect. Some configuration metrics had contradictory effects
on LST between the two cities. For example, after controlling for the ef-
fects of percent cover of trees, mean patch size was positively correlated
to LST in Baltimore, but was negatively correlated in Sacramento. Be-
cause larger patches have lower edge densities (Table A2), it follows
that edge density was negatively correlated to LST in Baltimore, but
was positively correlated in Sacramento. Previous studies on different
cities have also found contradictory results of spatial configuration of
greenspace/tree canopy on LST. For example, edge density of vegetation
cover was found to be negatively correlated with LST in Baltimore (Zhou
et al., 2011), Shanghai (Li et al., 2011; Li et al., 2014), and Berlin (Dugord
et al., 2014), but positive in Beijing (Li et al., 2013b). Our results from the
comparison of the two cities indicated that the spatial configuration of
trees may have different effects on LST in cities with different climatic
conditions. These results enhance the understanding of the inconsisten-
cy of effects of spatial configuration of trees/greenspace on LST from
previous studies.
Trees ameliorate temperatures primarily in two ways: providing
shade and through evapotranspiration. The contradictory results of con-
figuration metrics found in the two cities may be due to differences in
the relative contributions of the two cooling processes and these differ-
ences may be related to different climatic conditions between the cities.
Here, we again take edge density as an example. Increasing total edges
and edge density may potentially lead to an increase of shade provided
by trees to surrounding surfaces (Li et al., 2012; Zhou et al., 2011). In ad-
Fig. 2. The results of variance partitioning for percent cover of tree canopy and spatial dition, greater total edges and edge density may also enhance energy
configuration across spatial scales. flow and exchange between trees and their surrounding areas
 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12 7

Table 4
The results of spatial error models. The bold and italic rows are standardized coefficients.

City Scale PTree AREA_MN SHAPE_MN ED LPI R^2 AIC

Baltimore 120 m −0.059⁎⁎ 9.370E−06 −0.122⁎⁎ −1.376E−06 0.004⁎⁎ 0.932 42,232.200

−0.349 0.005 −0.012 −0.011 0.026
360 m −0.148⁎⁎ 2.432E−05⁎ −0.997⁎⁎ 1.072E−05⁎⁎ 0.010 0.900 4663.350
−0.790 0.025 −0.044 0.067 0.052
600 m −0.171⁎⁎ 7.819E−05⁎⁎ −5.565⁎⁎ 2.306E−05⁎⁎ 0.010 0.920 1405.970
−0.900 0.060 −0.130 0.137 0.048
840 m −0.165⁎⁎ 8.307E−05 −10.560⁎⁎ 2.135E−05⁎⁎ 0.001 0.924 641.347
−0.858 0.034 −0.207 0.123 0.004
1080 m −0.164⁎⁎ 5.917E−05 −11.100⁎⁎ 1.735E−05⁎⁎ −0.003 0.941 317.275
−0.846 0.035 −0.191 0.097 −0.013
Sacramento 120 m −0.063⁎⁎ 0.000 −0.251⁎⁎ 5.363E−06⁎⁎ 0.003 0.865 51,781.6
−0.319 0.001 −0.035 0.114 0.011
360 m −0.114⁎⁎ −0.007⁎⁎ −0.218 5.921E−06⁎ 0.015 0.741 5875.27
−0.572 −0.211 −0.013 0.120 0.035
600 m −0.047⁎ −0.015⁎⁎ 2.631⁎⁎ −8.548E-06 0.001 0.763 1827.85
−0.243 −0.480 0.102 −0.173 0.002
840 m −0.072⁎ −0.021⁎⁎ 7.121⁎⁎ −8.109E−06 0.061⁎ 0.827 744.035
−0.379 −0.637 0.193 −0.167 0.102
1080 m 0.019 −0.034⁎⁎ 8.371⁎⁎ −2.298E−05⁎⁎ 0.029 0.838 379.032
0.101 −0.932 0.217 −0.469 0.048
⁎⁎ P b 0.01.
⁎ P b 0.05 (2-tailed).

(Cadenasso et al., 2003; Zhou et al., 2011). Consequently, considering 2010; Yokohari et al., 1997; Zhang et al., 2009), but reduced shading ef-
only the shading process, increasing edge density will lead to lower fects. An increase in mean patch size will likely lead to reduced shading
LST. However, increased edge density is frequently a result of more effects because given a fixed amount of tree cover, an increase in mean
fragmented tree cover, given a fixed amount of total tree coverage. As patch size leads to a decrease in edge density (Table A2), which results
large and continuous tree stands generally have lower temperature in reduced shading effects, as discussed above. In the hotter and drier
than that of fragmented and smaller patches (Cao et al., 2010; Sacramento area, the increased evapotranspiration caused by increased
Yokohari et al., 1997; Zhang et al., 2009), suggesting stronger evapo- mean patch size is likely to outweigh reduction in shading. Therefore,
transpiration efficiency of larger patches, increasing edge density is like- mean patch size has a negative relationship with LST, given a fixed
ly to reduce evapotranspiration efficiency. This is particularly amount of tree coverage. In Baltimore, however, reduction in shading
predominant in cities such as Sacramento that have very dry and hot outweighed increased evapotranspiration, and thus an increase in
summers, during which vegetation is very likely to experience water mean patch size led to higher LST.
and temperature stress (Connors et al., 2013; Maimaitiyiming et al., Notably, the relative importance of mean patch size in predicting LST
2014). This is because the ambient temperature and humidity affect increased with the increased size of analytical unit in Sacramento, but
the transpiration rate of trees in a non-linear (an inverted U shape) the opposite was found in Baltimore, both suggesting clear scale effects.
way (Lambers et al., 2008; Schulze et al., 2005). That is, while increasing These scale effects may suggest that the two cooling processes, shading
temperature and reducing humidity to some extent can induce the sto- and evapotranspiration of trees, and their relative importance, change
mata open and thus enhance transpiration, excessive heat and increas- with scale, and differ by cities with different climatic conditions. This
ing vapor pressure deficit between leaf and air will lead to dramatic hypothesis, however, warrants further research.
reduction in transpiration (Lambers et al., 2008; Schulze et al., 2005). The relative importance of percent cover of trees, and spatial config-
Therefore, whether the increase of edge density will lead to a decrease uration on LST also varied greatly between the two cities. Percent cover
or increase in LST will largely depend on the net effects of increased of trees was the most important variable in predicting LST in Baltimore.
shading effects and reduced evapotranspiration effects. In Mediterra- This is consistent with many of the previous studies that have found
nean climate cities such as Sacramento, the reduction in evapotranspira- that percent cover of trees (or greenspaces) plays a more important
tion caused by increased edge density is likely to outweigh increased role than their spatial configuration (Li et al., 2012; Xie et al., 2013;
shading. Consequently, edge density has a positive relationship with Zhou et al., 2011). However, spatial configuration of tree cover, such
LST, given a fixed amount of tree coverage. But this is the opposite in as the mean patch size, played a more important role in predicting LST
cities such as Baltimore that experience a relative humid summer. than the percent cover of trees in Sacramento. In fact, the importance
Similar to edge density, whether the increase of mean patch size of percent cover of trees in predicting LST decreased with the increase
leads to a decrease or increase in LST largely depends on the joint effects of the size of analytical unit, and even became insignificant at the size
of the two key cooling processes, shading and evapotranspiration of of 840 m and greater (Table 3). This result is similar to the findings of
trees. In contrast to edge density, an increase in mean patch size will Maimaitiyiming et al. (2014) in a study conducted in Aksu, Xinjiang,
likely result in increased evapotranspiration efficiency (Cao et al., China, and of Li et al. (2016) in a study of Phoenix, Arizona, USA. Both
cities are relatively dry and hot in summer, similar to Sacramento.
These results indicated that the relative importance of percent cover
Table 5
of trees and their spatial configuration may vary by cities with different
Results from OLS linear regression. The response variable, LST, was predicted by PTree.
climatic conditions. It should be noted, however, that at the finest scale
Scale Baltimore Sacramento in this study – analytical unit of 120 m – percent cover of trees was a
Coef. R2 Coef. R2 much better predictor of LST than any configuration metrics in Sacra-
120 m −0.144 0.689 −0.129 0.409
mento (Table 3). With the recent availability of very fine resolution
360 m −0.173 0.817 −0.147 0.523 LST data (7 m resolution, e.g., Jenerette et al., 2016), research on how
600 m −0.18 0.858 −0.152 0.59 the relationship between spatial pattern of trees and LST varies by
840 m −0.184 0.877 −0.158 0.657 unit of analysis at a scale finer than 120 m would be highly desirable
1080 m −0.188 0.898 −0.16 0.671
to expand our understanding of the scale effects.
8 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12

4.2. The methodological implications: it is crucially important to choose the well as the approaches for scaling. Here, the spatial resolution of the
appropriate statistical approaches image data used to map tree cover was 1 m, but most previous studies
used the 30 m Landsat TM data.
Our results underscore the necessity of controlling for the effects of
percent cover of trees when quantifying the effects of spatial configura- 5. Conclusions
tion of tree cover on LST. For both cities, after controlling for the effects
of percent cover of trees (either through partial correlation or linear re- Urban greenspace, particularly trees, has significant cooling effects
gression modeling), the relationships between LST and configuration on urban heat. It is widely recognized that increasing percent coverage
metrics dramatically changed, compared with results from the Pearson of greenspace can greatly reduce ambient air temperatures and land
correlation analysis. For example, the relationship between LST and surface temperatures in urban environments. However, recent studies
mean patch size (AREA_MN) changed from negative to positive in Bal- investigating the effects of spatial configuration of greenspace show sig-
timore. Similarly, the relationship between LST and edge density (ED) nificant, but inconsistent results, including the direction of the effects.
in Sacramento changed from negative to positive. This is because most To investigate the causes of this inconsistency, we conducted a compar-
of the configuration variables are inherently correlated to percent ison study of Baltimore, MD and Sacramento, CA, USA, two cities with
cover of trees (Tables A3 & A4; Li and Wu, 2004; Peng et al., 2010; very different climatic conditions, using different statistical approaches
Riitters et al., 1995). For example, mean patch size had a significantly and analytical units with varied sizes. We found: (1) trees' cooling effi-
negative correlation with LST based on the Pearson correlation analysis ciency generally was higher in Baltimore than in the hotter and drier
(r = −0.56, P b 0.01; Table 2) in Baltimore at the scale of 120 m. This Sacramento. (2) The effects of spatial configuration of trees on LST var-
observed correlation, however, is due to the very strong positive corre- ied greatly in terms of magnitude, significance, and even direction, be-
lation between mean patch size and percent cover of trees (r = 0.70, P b tween the two cities, suggesting spatial configuration of trees may
0.01; Table A3). After controlling for the effect of percent cover of trees, play different roles in cities with different climatic conditions. Percent
mean patch size in fact had a significantly positive correlation with LST, cover of trees was more important than their spatial configuration in
due to the reasons we discussed in Section 4.1. Therefore, it is crucially predicting LST in Baltimore, but the opposite was found in Sacramento.
important to use statistical methods such as partial correlation and mul- Therefore, urban planners and managers should be cautious about di-
tiple regression models, instead of Pearson correlation, to assess the rel- rectly applying results found in cities with different climatic conditions.
ative contributions of percent cover of trees and configuration to LST. (3) When using different statistical approaches, the relationships be-
Using Pearson correlation analysis may generate misleading results. tween LST and configuration metrics could dramatically change. Our re-
Other statistical approaches, such as path analysis and structural sults underscore the necessity of controlling the effects of percent cover
equation modeling have been increasingly used to identify the complex of trees, when quantifying the effects of spatial configuration of trees on
and nested relationships among social conditions, land cover and sur- LST. These results contribute to the understanding of the inconsistent
face temperatures (Jenerette et al., 2007; Huang and Cadenasso, 2016; results from previous studies, which may be caused by the different
Tayyebi and Darrel Jenerette, 2016), which potentially allow the evalu- methods applied (e.g., Pearson correlation analysis versus partial corre-
ation of direct and indirect effects of tree cover and configuration on LST. lation). (4) Spatial autocorrelation could influence the relationships be-
Our results also showed that the spatial autocorrelation could influ- tween landscape metrics and LST, particularly when the unit of analysis
ence the relationships between landscape metrics and LST. This is par- is relatively small. (5) With the increase of the size of analytical unit, the
ticularly true when the unit of analysis is relatively small. However, relationships between spatial configuration metrics and LST became
when the unit of analysis in this study is relatively large (i.e., equal to stronger. This study can enhance the understanding on the effects of
or greater than a linear dimension of 600 m), results from OLS modeling spatial configuration of greenspace on UHI. It also provides important
and SAR modeling were similar, in terms of both regression coefficient insights to urban planners and natural resource managers on how to
and R2. This may suggest that the frequently used OLS is appropriate mitigate the impact of urbanization on UHI through urban design and
at such scales. vegetation management.
We found that with increasing size of the analytical unit, the rela-
tionships between LST and spatial pattern metrics, including both per- Acknowledgments
cent cover and configuration, became stronger. The spatial pattern of
tree cover also explained more variation in LST. We did not find a The support of the National Natural Science Foundation of China
“best” size of analytical unit, at which the correlations (or R2) peaked, (41422104 and 41371197) and the One Hundred Talents program
and a turning point occurred (Liu and Weng, 2009; Peng et al., 2016; of Chinese Academy of Sciences (29BR2013002) is gratefully
Weng et al., 2004). This may be due to the very different data used, as acknowledged.

Appendix A. Appendix

Table A1
A descriptive statistics of LST and landscape metrics of trees.

City Scale LST PTree AREA_MN SHAPE_MN ED LPI

Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

Baltimore 120 31.87 4.14 27.51 24.45 950.02 2267.63 1.37 0.40 57,518.11 33,468.78 19.25 23.91
360 31.87 3.83 28.11 20.52 1191.00 3949.92 1.37 0.17 52,488.62 24,066.84 14.66 19.97
600 31.87 3.56 28.13 18.73 980.78 2740.28 1.35 0.08 51,759.32 21,151.46 12.34 17.99
840 31.78 3.40 28.83 17.71 866.37 1379.75 1.35 0.07 52,364.87 19,491.47 11.50 17.04
1080 31.68 3.28 29.73 16.89 901.85 1933.74 1.34 0.06 53,401.59 18,325.08 11.26 16.30
city 31.87 4.14 27.10 599.60 19,526.35 1.32 0.54 478.26 2.14
Sacramento 120 33.27 2.97 16.93 15.10 81.65 243.84 1.25 0.41 90,008.97 62,935.26 6.82 9.99
360 33.28 2.56 17.29 12.87 79.54 81.49 1.32 0.15 88,252.77 51,997.52 3.43 6.04
600 33.31 2.35 17.42 12.12 80.66 73.43 1.33 0.09 87,813.52 47,579.94 2.39 4.94
840 33.35 2.18 17.41 11.49 80.57 67.24 1.32 0.06 87,552.63 44,935.50 1.87 3.66
1080 33.30 2.06 18.00 10.81 81.28 57.09 1.33 0.05 89,677.74 42,027.10 1.72 3.43
city 33.27 2.97 16.66 73.80 1098.00 1.39 3.33 819.85 0.03
 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12 9

Table A2
Partial correlation between mean patch size and edge density controlling for the effect of percent cover of trees.

Baltimore Sacramento

120 −0.655** −0.438**

360 −0.475** −0.746**
600 −0.454** −0.789**
840 −0.643** −0.807**
1080 −0.528** −0.846**

**P b 0.01 (2-tailed).

Table A3
Correlation matrix between independent variables in Baltimore.

Scale PTree AREA_MN SHAPE_MN ED LPI

120 m PTree 1
AREA_MN 0.70** 1
SHAPE_MN 0.51** 0.31** 1
ED 0.66** 0.11** 0.54** 1
LPI 0.95** 0.78** 0.45** 0.45** 1
360 m PTree 1
AREA_MN 0.53** 1
SHAPE_MN 0.61** 0.46** 1
ED 0.61** 0.00 0.44** 1
LPI 0.91** 0.61** 0.52** 0.31** 1
600 m PTree 1
AREA_MN 0.54** 1
SHAPE_MN 0.73** 0.50** 1
ED 0.62** 0.04* 0.53** 1
LPI 0.88** 0.64** 0.52** 0.29** 1
840 m PTree 1
AREA_MN 0.76** 1
SHAPE_MN 0.72** 0.46** 1
ED 0.63** 0.16 0.65** 1
LPI 0.87** 0.87** 0.45** 0.27** 1
1080 m PTree 1
AREA_MN 0.60** 1
SHAPE_MN 0.72** 0.34** 1
ED 0.62** 0.05 0.67** 1
LPI 0.86** 0.71** 0.40** 0.25** 1

**P b 0.01, *P b 0.05 (2-tailed).

Table A4
Correlation matrix between independent variables in Sacramento.

Scale PTree AREA_MN SHAPE_MN ED LPI

120 m Ptree 1
AREA_MN 0.40** 1
SHAPE_MN 0.44** 0.20** 1
ED 0.85** 0.12** 0.48** 1
LPI 0.84** 0.52** 0.31** 0.50** 1
360 m PTree 1
AREA_MN 0.67** 1
SHAPE_MN 0.41** 0.38** 1
ED 0.87** 0.31** 0.40** 1
LPI 0.70** 0.76** 0.19** 0.36** 1
600 m PTree 1
AREA_MN 0.69** 1
SHAPE_MN 0.48** 0.46** 1
ED 0.87** 0.32** 0.44** 1
LPI 0.61** 0.72** 0.14** 0.28** 1
840 m PTree 1
AREA_MN 0.70** 1
SHAPE_MN 0.69** 0.57** 1
ED 0.88** 0.34** 0.65** 1
LPI 0.61** 0.75** 0.22** 0.28** 1
1080 m PTree 1
AREA_MN 0.73** 1
SHAPE_MN 0.69** 0.62** 1
ED 0.88** 0.38** 0.64** 1
LPI 0.57** 0.69** 0.21* 0.27** 1

**P b 0.01, *P b 0.05 (2-tailed).

10 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12

Fig. A1. Scattergrams of land surface temperature (LST) versus·percent cover of tree canopy across all scales at two cities: B120, B360, B600, B840 and B1080: Baltimore; S120, S360, S600, S800 and
S1080: Sacramento.
 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12 11

References
Akbari, H., Rosenfeld, A., Taha, H., Gartland, L., 1996. Mitigation of summer urban heat Lee, S.W., Hwang, S.J., Lee, S.B., Hwang, H.S., Sung, H.C., 2009. Landscape ecological ap-
islands to save electricity and smog. 76th Annual Meteorological Society Meeting. proach to the relationships of land use patterns in watersheds to water quality char-
GA, Atlanta. acteristics. Landsc. Urban Plan. 92, 80–89.
Akbari, H., Pomerantz, M., Taha, H., 2001. Cool surfaces and shade trees to reduce energy Li, H.B., Wu, J.G., 2004. Use and misuse of landscape indices. Landsc. Ecol. 19, 389–399.
use and improve air quality in urban areas. Sol. Energy 70, 295–310. Li, S.C., Zhao, Z.Q., Xie, M.M., Wang, Y.L., 2010. Investigating spatial non-stationary and
Anderson, M.J., Gribble, N.A., 1998. Partitioning the variation among spatial, temporal and scale-dependent relationships between urban surface temperature and environmen-
environmental components in a multivariate data set. Aust. J. Ecol. 23, 158–167. tal factors using geographically weighted regression. Environ. Model. Softw. 25,
Anselin, L., 2005a. Exploring Spatial Data With GeoDa™: A Workbook. University of Illi- 1789–1800.
nois, Urbana. Li, J.X., Song, C.H., Cao, L., Zhu, F.G., Meng, X.L., Wu, J.G., 2011. Impacts of landscape struc-
Anselin, L., 2005b. Spatial Regression Analysis in R: A Workbook. University of Illinois, ture on surface urban heat islands: a case study of Shanghai, China. Remote Sens. En-
Urbana. viron. 115, 3249–3263.
Arnfield, A.J., 2003. Two decades of urban climate research: a review of turbulence, ex- Li, X.M., Zhou, W.Q., Ouyang, Z.Y., Xu, W.H., Zheng, H., 2012. Spatial pattern of greenspace
changes of energy and water, and the urban heat island. Int. J. Climatol. 23, 1–26. affects land surface temperature: evidence from the heavily urbanized Beijing metro-
Asgarian, A., Amiri, B.J., Sakieh, Y., 2015. Assessing the effect of green cover spatial pat- politan area, China. Landsc. Ecol. 27, 887–898.
terns on urban land surface temperature using landscape metrics approach. Urban Li, X., Zhou, W., Ouyang, Z., 2013a. Forty years of urban expansion in Beijing: what is the
Ecosystems 18, 209–222. relative importance of physical, socioeconomic, and neighborhood factors? Appl.
Barsi, J.A., Schott, J.R., Palluconi, F.D., Hook, S.J., 2005. Validation of a web-based atmo- Geogr. 38, 1–10.
spheric correction tool for single thermal band instruments. Optics & Photonics Li, X.M., Zhou, W.Q., Ouyang, Z.Y., 2013b. Relationship between land surface temperature
2005. International Society for Optics and Photonics (pp. 58820E-58820E-58827). and spatial pattern of greenspace: what are the effects of spatial resolution? Landsc.
Bowler, D.E., Buyung-Ali, L., Knight, T.M., Pullin, A.S., 2010. Urban greening to cool towns Urban Plan. 114, 1–8.
and cities: a systematic review of the empirical evidence. Landsc. Urban Plan. 97, Li, W.F., Bai, Y., Chen, Q.W., He, K., Ji, X.H., Han, C.M., 2014. Discrepant impacts of land use
147–155. and land cover on urban heat islands: a case study of Shanghai, China. Ecol. Indic. 47,
Brazel, A., Selover, N., Vose, R., Heisler, G., 2000. The tale of two climates—Baltimore and 171–178.
Phoenix urban LTER sites. Clim. Res. 15, 123–135. Li, X.X., Li, W.W., Middel, A., Harlan, S., Brazel, A., Turner, B., 2016. Remote sensing of the
Buyantuyev, A., Wu, J.G., 2010. Urban heat islands and landscape heterogeneity: linking surface urban heat island and land architecture in Phoenix, Arizona: combined effects
spatiotemporal variations in surface temperatures to land-cover and socioeconomic of land composition and configuration and cadastral–demographic–economic factors.
patterns. Landsc. Ecol. 25, 17–33. Remote Sens. Environ. 174, 233–243.
Cadenasso, M.L., Pickett, S.T., Weathers, K.C., Jones, C.G., 2003. A framework for a theory of Lichstein, J.W., Simons, T.R., Shriner, S.A., Franzreb, K.E., 2002. Spatial autocorrelation and
ecological boundaries. Bioscience 53, 750–758. autoregressive models in ecology. Ecol. Monogr. 72, 445–463.
Cao, X., Onishi, A., Chen, J., Imura, H., 2010. Quantifying the cool island intensity of urban Liu, H., Weng, Q.H., 2009. Scaling effect on the relationship between landscape pattern
parks using ASTER and IKONOS data. Landsc. Urban Plan. 96, 224–231. and land surface temperature: a case study of Indianapolis, United States.
Chander, G., Markham, B., 2003. Revised Landsat-5 TM radiometric calibration procedures Photogramm. Eng. Remote. Sens. 75, 291–304.
and postcalibration dynamic ranges. IEEE Trans. Geosci. Remote Sens. 41, 2674–2677. Ma, Y., Kuang, Y.Q., Huang, N.S., 2010. Coupling urbanization analyses for studying urban
Chander, G., Markham, B.L., Helder, D.L., 2009. Summary of current radiometric calibra- thermal environment and its interplay with biophysical parameters based on TM/
tion coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote Sens. En- ETM plus imagery. Int. J. Appl. Earth Obs. Geoinf. 12, 110–118.
viron. 113, 893–903. MacFaden, S.W., O'Neil-Dunne, J.P.M., Royar, A.R., Lu, J.W.T., Rundle, A.G., 2012. High-res-
Chen, A.L., Yao, X.A., Sun, R.H., Chen, L.D., 2014. Effect of urban green patterns on surface olution tree canopy mapping for New York City using LIDAR and object-based image
urban cool islands and its seasonal variations. Urban For. Urban Green. 13, 646–654. analysis. J. Appl. Remote. Sens. 6, 1–23.
Connors, J.P., Galletti, C.S., Chow, W.T.L., 2013. Landscape configuration and urban heat is- Maimaitiyiming, M., Ghulam, A., Tiyip, T., Pla, F., Latorre-Carmona, P., Halik, U., Sawut, M.,
land effects: assessing the relationship between landscape characteristics and land Caetano, M., 2014. Effects of green space spatial pattern on land surface temperature:
surface temperature in Phoenix, Arizona. Landsc. Ecol. 28, 271–283. implications for sustainable urban planning and climate change adaptation. ISPRS
Dugord, P.-A., Lauf, S., Schuster, C., Kleinschmit, B., 2014. Land use patterns, temperature J. Photogramm. Remote Sens. 89, 59–66.
distribution, and potential heat stress risk – the case study Berlin, Germany. Comput. McCarthy, H.R., Pataki, D.E., Jenerette, G.D., 2011. Plant water-use efficiency as a metric of
Environ. Urban. Syst. 48, 86–98. urban ecosystem services. Ecol. Appl. 21, 3115–3127.
Fan, C., Myint, S.W., Zheng, B.J., 2015. Measuring the spatial arrangement of urban vege- McGarigal, K., Cushman, S., Neel, M., Ene, E., 2002. FRAGSTATS: spatial pattern analysis
tation and its impacts on seasonal surface temperatures. Prog. Phys. Geogr. 39, program for categorical maps. Computer Software Program Produced by the Authors
199–219. at the University of Massachusetts, Amherst (Available at the following web site:).
Fouillet, A., Rey, G., Laurent, F., Pavillon, G., Bellec, S., Guihenneuc-Jouyaux, C., Clavel, J., http://www.umass.edu/landeco/research/fragstats/fragstats.
Jougla, E., Hémon, D., 2006. Excess mortality related to the August 2003 heat wave Myint, S.W., Zheng, B.J., Talen, E., Fan, C., Kaplan, S., Middel, A., Smith, M., Huang, H.-P.,
in France. Int. Arch. Occup. Environ. Health 80, 16–24. Brazel, A., 2015. Does the spatial arrangement of urban landscape matter? Examples
Gustafson, E.J., 1998. Quantifying landscape spatial pattern: what is the state of the art? of urban warming and cooling in Phoenix and Las Vegas. Ecosystem Health Sustain. 1,
Ecosystems 1, 143–156. 1–15.
Hamada, S., Ohta, T., 2010. Seasonal variations in the cooling effect of urban green areas Niemelä, J., 1999. Ecology and urban planning. Biodivers. Conserv. 8, 119–131.
on surrounding urban areas. Urban For. Urban Green. 9, 15–24. Pataki, D.E., McCarthy, H.R., Litvak, E., Pincetl, S., 2011. Transpiration of urban forests in
Harlan, S.L., Ruddell, D.M., 2011. Climate change and health in cities: impacts of heat and the Los Angeles metropolitan area. Ecol. Appl. 21, 661–677.
air pollution and potential co-benefits from mitigation and adaptation. Curr. Opin. Peng, J., Wang, Y.L., Zhang, Y., Wu, J.S., Li, W.F., Li, Y., 2010. Evaluating the effectiveness of
Environ. Sustain. 3, 126–134. landscape metrics in quantifying spatial patterns. Ecol. Indic. 10, 217–223.
Heikkinen, R.K., Luoto, M., Kuussaari, M., Poyry, J., 2005. New insights into butterfly- Peng, J., Xie, P., Liu, Y.X., Ma, J., 2016. Urban thermal environment dynamics and associat-
environment relationships using partitioning methods. Proc. Biol. Sci. 272, ed landscape pattern factors: a case study in the Beijing metropolitan region. Remote
2203–2210. Sens. Environ. 173, 145–155.
Huang, G., Cadenasso, M.L., 2016. People, landscape, and urban heat island: dynamics Polsky, C., Grove, J.M., Knudson, C., Groffman, P.M., Bettez, N., Cavender-Bares, J., Hall, S.J.,
among neighborhood social conditions, land cover and surface temperatures. Landsc. Heffernan, J.B., Hobbie, S.E., Larson, K.L., 2014. Assessing the homogenization of urban
Ecol. 31, 2507–2515. land management with an application to US residential lawn care. Proc. Natl. Acad.
Huang, G.L., Zhou, W.Q., Cadenasso, M.L., 2011. Is everyone hot in the city? Spatial pattern Sci. 111, 4432–4437.
of land surface temperatures, land cover and neighborhood socioeconomic character- Poumadere, M., Mays, C., Le Mer, S., Blong, R., 2005. The 2003 heat wave in France: dan-
istics in Baltimore, MD. J. Environ. Manag. 92, 1753–1759. gerous climate change here and now. Risk Anal. 25, 1483–1494.
Imhoff, M.L., Zhang, P., Wolfe, R.E., Bounoua, L., 2010. Remote sensing of the urban heat R Development Core Team, 2011. R: A Language and Environment for Statistical Comput-
island effect across biomes in the continental USA. Remote Sens. Environ. 114, ing. Version 2.12.1R Foundation for Statistical Computing, Vienna.
504–513. R. Team, 2011. A Language and Environment for Statistical Computing. R Development
Jenerette, G.D., Harlan, S.L., Brazel, A., Jones, N., Larsen, L., Stefanov, W.L., 2007. Regional Core Team.
relationships between surface temperature, vegetation, and human settlement in a Rhee, J., Park, S., Lu, Z., 2014. Relationship between land cover patterns and surface tem-
rapidly urbanizing ecosystem. Landsc. Ecol. 22, 353–365. perature in urban areas. Geosci. Remote Sens. Lett. 51, 521–536.
Jenerette, G.D., Harlan, S.L., Buyantuev, A., Stefanov, W.L., Declet-Barreto, J., Ruddell, B.L., Riitters, K.H., O'Neill, R.V., Hunsaker, C.T., Wickham, J.D., Yankee, D.H., Timmins, S.P., Jones,
Myint, S.W., Kaplan, S., Li, X., 2016. Micro-scale urban surface temperatures are relat- K.B., Jackson, B.L., 1995. A factor analysis of landscape pattern and structure metrics.
ed to land-cover features and residential heat related health impacts in Phoenix, AZ Landsc. Ecol. 10, 23–39.
USA. Landsc. Ecol. 31, 745–760. Santamouris, M., Cartalis, C., Synnefa, A., Kolokotsa, D., 2015. On the impact of urban heat
Kong, F.H., Yin, H.W., Wang, C.Z., Cavan, G., James, P., 2014. A satellite image-based anal- island and global warming on the power demand and electricity consumption of
ysis of factors contributing to the green-space cool island intensity on a city scale. buildings—a review. Energ. Buildings 98, 119–124.
Urban For. Urban Green. 13, 846–853. Schulze, E.-D., Beck, E., Müller-Hohenstein, K., 2005. Plant Ecology. Springer, Berlin,
Lambers, H., Chapin III, F.S., Pons, T.L., 2008. Plant Physiological Ecology. Springer, New pp. 253–379.
York, pp. 163–217. Sobrino, J.A., Jiménez-Muñoz, J.C., Paolini, L., 2004. Land surface temperature retrieval
Landsat Project Science Office, 2009. Landsat 7 Science Data Users Handbook. Goddard from LANDSAT TM 5. Remote Sens. Environ. 90, 434–440.
Space Flight Center, NASA, Washington, DC (URL: http://landsathandbook.gsfc.nasa. Sun, R.H., Chen, L.D., 2017. Effects of green space dynamics on urban heat islands: Mitiga-
gov/handbook.html). tion and diversification. Ecosyst. Serv. 23, 38–46.
12 W. Zhou et al. / Remote Sensing of Environment 195 (2017) 1–12

Tayyebi, A., Darrel Jenerette, G., 2016. Increases in the climate change adaption effective- Yan, H., Fan, S.X., Guo, C.X., Wu, F., Zhang, N., Dong, L., 2014. Assessing the effects of land-
ness and availability of vegetation across a coastal to desert climate gradient in met- scape design parameters on intra-urban air temperature variability: the case of Bei-
ropolitan Los Angeles, CA, USA. Sci. Total Environ. 548–549, 60–71. jing, China. Build. Environ. 76, 44–53.
Voogt, J.A., Oke, T.R., 2003. Thermal remote sensing of urban climates. Remote Sens. Envi- Yokohari, M., Brown, R.D., Kato, Y., Moriyama, H., 1997. Effects of paddy fields on sum-
ron. 86, 370–384. mertime air and surface temperatures in urban fringe areas of Tokyo, Japan. Landsc.
Wan, K.K.W., Li, D.H.W., Pan, W.Y., Lam, J.C., 2012. Impact of climate change on building Urban Plan. 38, 1–11.
energy use in different climate zones and mitigation and adaptation implications. Yuan, F., Bauer, M.E., 2007. Comparison of impervious surface area and normalized differ-
Appl. Energy 97, 274–282. ence vegetation index as indicators of surface urban heat island effects in Landsat im-
Wang, H., Ouyang, Z., Chen, W., Wang, X., Zheng, H., Ren, Y., 2011. Water, heat, and air- agery. Remote Sens. Environ. 106, 375–386.
borne pollutants effects on transpiration of urban trees. Environ. Pollut. 159, Zhang, X.Y., Zhong, T.Y., Feng, X.Z., Wang, K., 2009. Estimation of the relationship between
2127–2137. vegetation patches and urban land surface temperature with remote sensing. Int.
Weng, Q.H., 2009. Thermal infrared remote sensing for urban climate and environmental J. Remote Sens. 30, 2105–2118.
studies: methods, applications, and trends. ISPRS J. Photogramm. Remote Sens. 64, Zhou, W.Q., Troy, A., 2008. An object-oriented approach for analysing and characterizing
335–344. urban landscape at the parcel level. Int. J. Remote Sens. 29, 3119–3135.
Weng, Q.H., Lu, D.S., Schubring, J., 2004. Estimation of land surface temperature–vegeta- Zhou, W.Q., Troy, A., Grove, M., 2008. Modeling residential lawn fertilization practices: in-
tion abundance relationship for urban heat island studies. Remote Sens. Environ. tegrating high resolution remote sensing with socioeconomic data. Environ. Manag.
89, 467–483. 41, 742–752.
Weng, Q.H., Lu, D.S., Liang, B.Q., 2006. Urban surface biophysical descriptors and land sur- Zhou, W.Q., Huang, G.L., Cadenasso, M.L., 2011. Does spatial configuration matter? Under-
face temperature variations. Photogramm. Eng. Remote. Sens. 72, 1275–1286. standing the effects of land cover pattern on land surface temperature in urban land-
White, M.A., Nemani, R.R., Thornton, P.E., Running, S.W., 2002. Satellite evidence of phe- scapes. Landsc. Urban Plan. 102, 54–63.
nological differences between urbanized and rural areas of the eastern United States Zhou, W.Q., Qian, Y.G., Li, X.M., Li, W.F., Han, L.J., 2014. Relationships between land cover
deciduous broadleaf forest. Ecosystems 5, 260–273. and the surface urban heat island: seasonal variability and effects of spatial and the-
Wu, H., Ye, L.P., Shi, W.Z., Clarke, K.C., 2014. Assessing the effects of land use spatial struc- matic resolution of land cover data on predicting land surface temperatures. Landsc.
ture on urban heat islands using HJ-1B remote sensing imagery in Wuhan, China. Int. Ecol. 29, 153–167.
J. Appl. Earth Obs. Geoinf. 32, 67–78. Zhou, W.Q., Pickett, S.T.A., Cadenasso, M.L., 2016. Shifting concepts of urban spatial
Xie, M.M., Wang, Y.L., Chang, Q., Fu, M.C., Ye, M.T., 2013. Assessment of landscape patterns heterogeneity and their implications for sustainability. Landsc. Ecol. 1–16.
affecting land surface temperature in different biophysical gradients in Shenzhen,
China. Urban Ecosystems 16, 871–886.